【PPLR】Part-based Pseudo Label Refinement for Unsupervised Person Re-identification

论文信息
- 论文名：Part-based Pseudo Label Refinement for Unsupervised Person Re-identification
- 领域：unsupervised person reid
- 关键词：part-based, fully unsupervised, 可以用/不用额外的camera ids
- 作者/机构：KAIST
- 发表年份：2022
- 会议/期刊名：CVPR
- 级别：CCF-A
- 论文地址：Part-Based Pseudo Label Refinement for Unsupervised Person Re-Identification (thecvf.com)
- 项目地址：yoonkicho/PPLR: [CVPR 2022] Part-based Pseudo Label Refinement for Unsupervised Person Re-identification (github.com)
ABSTRACT & INTRODUCTION
- 目前存在的问题
  - Pseudo-labels are inherently noisy and deteriorate the accuracy
  - 解决思路：reduces the label noise by employing the complementary relationship between global and part features
- 本文方法和创新点
  - ① A cross agreement score as the similarity of k-nearest neighbors between feature spaces to exploit the reliable complementary relationship
  - ② Refine global/part features
    
    Refine pseudo-labels of global features by ensembling the predictions of part features
    
    Refine pseudo-labels of part features by applying label smoothing according to the suitability of given labels for each part
- 实验效果提升
  - Market1501, MSMT new SOTA
RELATED WORK
- Person ReID
  - Problem: require a large amount of labeled data that demands costly annotations
  - Solution: unsupervised
- Prior works on unsupervised ReID
PROPOSED METHOD
- 模型结构
  - Clustering stage
  - Training stage
- 模型分解—1 Part-based unsupervised re-ID framework --> Baseline
  - Losses
    
    Global classification cross-entropy loss
    
    Part classification cross-entropy loss
    
    Softmax-triplet loss
    
    Inter-camera contrastive loss (optional如果可以得到camera ids)
- 模型分解—2 Cross agreement
- 模型分解—3 Pseudo label refinement
  - Refine part labels—Agreement-aware label smoothing
    
    Label smoothing
    
    改写后的part classification loss
  - Refine global labels—Part-guided label refinement
    
    利用part pseudo-labels refine后的global pseudo-label
    
    Ensemble weight
    
    Weighting parameter controllingthe ratio of the one-hot pseudo-label and the ensembled prediction： β
    
    改写后的global classification loss
  - 改写以上两部分后的overall training objective
EXPERIMENT
- Implementation details
  - Input size: 384*128
  - Batchsize: 64
  - Clustering algorithm: DBSCAN
- 数据集
  
  数据来源链接
  - ① Market-1501-v15.09.15
  - ② MSMT17_V1
  - ③ VeRi
- 对比算法
  
  源码链接
  - ①
  - ②
  - ③
- 结果
  - Comparison with State-of-the-art Methods
总结和思考( •̀ ω •́ )✧
- ①
- ②
- ③